Projects

An open source alternative to ManyChat: automates Instagram comment-to-DM replies and surfaces account analytics, self-hosted with your own API keys so you own the data and the Meta app. Built to be forked and deployed in minutes rather than rented per seat. Reached 300+ GitHub stars and 1M+ views on Instagram.

My personal blog. Goal is to familiarize with the Instagram algorithm

Padel Court Keypoint Detection

Heatmap-based network with classical computer vision refinement

Adapted and trained a heatmap-based Modified TrackNet for keypoint detection in broadcast padel videos, achieving 95.09% validation accuracy (@ 7px) on a ~38,200 frame dataset (245 epochs). Deployed via ONNX with memory arena optimizations. Integrates auxiliary supervision & transfer learning for training stability, alongside homography matrix tracking and sympy line-intersections to extrapolate grid nodes.

Clutch RAG Shot Search

AI-powered semantic video search & coaching recommendation engine

Semantic video search app using a RAG pipeline & GPT-4o-mini to search padel video content. Generates personalized coaching recommendations based on player stats and matched clips. Features YouTube transcript ingestion and automated video segmentation for automated clip gathering.

Real-time Gaze Tracking System

Pilot project for Cluely — Privacy-first webcam gaze estimation

Built a lightweight, local-only gaze tracking prototype using standard webcam to infer eye/pupil position and head pose. The system classifies gaze direction (left/right/up/down/center) and emits continuous gaze vectors at 15–30 FPS. Features include swappable face/landmark detection, optional five-point calibration, temporal smoothing with confidence scores, and dual outputs: debug overlay with face/eye ROI visualization and programmatic stream (NDJSON console + optional WebSocket). Privacy-first architecture with no frame storage or network transmission, CPU-only performance with less than 100ms latency, and simple integration for immediate live gaze event streaming.

Steve the Freakysaur

Daydream Game Jam 2025 — 1st Place Winner

Created "Steve the Freakysaur," a Chrome Dino–style game controlled entirely hands-free using tongue detection through real-time facial landmark segmentation. Built during BC's biggest youth hackathon (Daydream Game Jam), which was also the largest Daydream event in North America, the project won 1st place out of 124 participants and 50+ projects. The game features single-player, two-player, and Flappy Bird–style modes, and was prototyped in Pygame before being ported to Next.js with TypeScript for web play.